Essential Math for AI

Hala Nelson, Essential Math for AI: Next-Level Mathematics for Efficient and Successful AI Systems (O'Reilly, 2023; ISBN 978-1-098-10763-5). Arrived as an Amazon link; body not read: this page rests on the publisher's description and the chapter topics, with bibliographic facts confirmed against OpenLibrary. The subject is mathematics, so it routes here rather than to ../machine-learning-wiki; the AI framing is the field the examples come from.

What it is, and why it is this spoke's

The book surveys the mathematics an AI practitioner needs (regression, neural networks and convolution, optimization, probability, graphs and random walks, Markov processes, and differential equations) and organizes the whole of it around AI applications: computer vision, natural language processing, generative models, reinforcement learning, and operations research. Its own pitch is the tell: it presents this "rather than dense academic theory," in "an immersive and conversational style," for an audience of engineers, data scientists and career-changers.

That places it cleanly on this spoke's side of its own boundary and tests the boundary at the same time. The rule is the maths is the subject here; the application is the subject there. Nelson's book teaches the maths, several bodies of it, and uses AI as the motivating field throughout. It does not advance a claim about a model; it teaches the linear algebra, calculus and probability a model rests on. So it routes here even though every example is an AI one, exactly as thibos-fourier-analysis (a vision-science methods course that is ~85% general Fourier analysis) did.

Where it sits on the corpus's axes

• Organized by application, not by mathematical structure. Every other text here is filed under a mathematical subject (linear-algebra, probability-theory, graph-theory) and teaches it in the order the mathematics suggests. Nelson selects and sequences the mathematics by what an AI system needs, which is a different organizing principle and the corpus's first instance of it. A chapter here is "the math for convolution," not "a chapter of analysis."
• Near the rigour floor, and openly so. "Rather than dense academic theory" puts it below the proof-complete texts and beside the corpus's applied end: corral-vector-calculus (author-rated 5/10) and illowsky-introductory-statistics (reaches its whole syllabus with no proof). A survey that touches regression through differential equations in one volume cannot derive any of them, and does not claim to.
• A sibling to the one text that argues from AI. edelman-johnson-matrix-calculus is Matrix Calculus (for Machine Learning and Beyond) and argues the calculus canon should be extended because linear algebra "has taken over" ML's tools. Nelson makes the same move at survey breadth rather than in one subject: the shape of the mathematics is chosen by what AI now needs of it.

Tier — T2, and the honest limits

T2: a professionally published survey by a working mathematician, specific and broad, but explicitly a bridge away from theory rather than a first-principles development, and its body was not read here — the assessment is of what the book is for, taken from the publisher's own description, not of how it derives anything. Two things a later reading should settle: the actual depth per subject (a survey this wide is necessarily thin somewhere), and whether it derives or merely states. The one page-count signal seen (an Amazon "1045 pages" field) is not trusted and is left off until confirmed.

Related

mathematics · linear-algebra · probability-theory · graph-theory · vector-calculus · matrix-calculus · edelman-johnson-matrix-calculus · corral-vector-calculus · illowsky-introductory-statistics · hala-nelson · oreilly-media · synthesis